- Rating
- 2
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper examines an important aspect of class-incremental learning: not only retaining previously learned knowledge, but also maintaining the ability to learn new classes. The authors investigate how unsupported outputs (classes not yet seen during training), can negatively affect learning when local classifier outputs are used as representations for subsequent layers. They show that these outputs participate in local softmax competition and receive gradient updates despite having no positive training examples, potentially impairing the acquisition of new classes. To address this issue, the paper proposes Anchor-Forward, a method that preserves forward competition while blocking direct gradients through unsupported outputs. The authors further analyze how the number of unsupported outputs, competition strength, layer depth, and exposure timing influence learning performance. Experiments on CIFAR-100 and GTSRB using DeeperForward and HCL-FF show that the proposed intervention can substantially recover new-task learning performance.
Critical strengths
The paper addresses an important yet relatively underexplored aspect of continual learning by focusing on the ability to acquire new knowledge rather than solely on preventing forgetting. The proposed Anchor-Forward intervention is simple yet effective, showing substantial improvements in new-task accuracy across different architectures and datasets.
Critical weaknesses
-
The negative effects of unsupported outputs could be easily avoided by incrementally expanding the classifiers or simply masking unseen classes during training. This raises concerns about whether the problem reflects realistic learning scenarios or is largely an artifact of the experimental setup. As a result, the paper seems to artificially introduce an otherwise avoidable problem and then address it with an unnecessarily complicated intervention, raising fundamental questions about the practical relevance and necessity of the proposed method.
-
The experiments assume a predefined output capacity, which requires knowing the total number of classes in advance. It would be much more practical to consider dynamic classifier expansion, where the output layer grows as new classes arrive.
-
The proposed Anchor-Forward consistently performs worse than the simpler M0 baseline in new-task accuracy. This raises questions about its practical value beyond serving as an analytical intervention, as the justification for using the proposed method over a simpler and better-performing alternative remains unclear.
-
The provided supplementary code appears to implement a substantially different method from the one described in the paper. Specifically, it does not include the proposed stop-gradient mechanism or the unsupported-class setting, making it difficult to reproduce the main experiments.
-
The evaluation is limited to relatively small datasets and architectures, and even CIFAR-100 is evaluated using only a subset of its training samples. Experiments with larger datasets and more realistic training settings would strengthen the findings.
-
Although the effect may be more pronounced in FF architectures that propagate intermediate classifier outputs, unsupported-output gradients are not inherently specific to FF. The appendix includes conventional backpropagation experiments, but these are too limited to establish whether the findings generalize to broader learning settings.
-
The stronger negative effects observed in shallow layers may depend on the specific architecture and learning mechanism of FF models. It remains unclear whether the same behavior would occur in networks trained with standard backpropagation.
-
The paper lacks sufficient conceptual discussion and experimental comparisons with prior work. Existing studies have reported similar issues with negative supervision [1] and even contrasting findings on the benefits of pre-allocated future-class outputs [2], but these connections are not adequately discussed. More concrete comparisons with related methods like [3–5] are needed to establish the novelty and generality of the proposed approach.
[1] Temporal Imbalance of Positive and Negative Supervision in Class-Incremental Learning
[2] Class-incremental Learning with Pre-allocated Fixed Classifiers
[3] A Contrastive Symmetric Forward-Forward Algorithm for Continual Learning Tasks
[4] SS-IL: Separated Softmax for Incremental Learning
[5] Loss of Plasticity in Deep Continual Learning
Questions for authors
The authors should address the concerns raised above, particularly by clarifying the practical relevance and justification of the proposed experimental setting and method. Given that simpler alternatives can avoid the identified problem, it is important to explain why Anchor-Forward is necessary and under what conditions it offers meaningful benefits. Furthermore, broader evaluations using standard backpropagation and more diverse tasks and architectures are essential to establish the generalizability and significance of the findings.
6 comments
Response 2: W3, W4 — M0 comparison and updated complete code
W3 | Comparison with M0
M0 has higher observed new-task means in all three comparisons. Original two-task current accuracy is 44.84% for Anchor and 46.40% for M0. The common-backbone two-task values are 43.60% and 44.72%. Full-pool mean arriving accuracy is 71.033% and 71.858%, respectively. These comparisons measure acquisition separately from final accuracy. Anchor tests whether acquisition can recover while unsupported coordinates remain in numerical normalization. Relative to Matched, the original and full-pool gains are 27.66 and 18.451 percentage points. Removing direct unsupported derivatives therefore recovers much of the acquisition loss while retaining denominator coordinates.
Changes made: Compared the acquisition means of Anchor and M0 directly (Sections 5.3, 5.6–5.7 and 6; Appendix I.8).
W4 | Code and reproducibility
The updated supporting materials contain complete source code for our method implementations in DeeperForward and HCL-FF. In each implementation, M0 restricts local normalization to supported classes, Matched includes the selected unsupported classes, and Anchor detaches only their loss-side scores. Class-associated feature maps still propagate. DeeperForward uses native sequential local updates; HCL-FF retains its local contrastive objective, projection heads, routed cross-entropy, and causal terminal classifier. The README identifies these distinct call chains and their readouts.
The implementation distinction is explicit in the native call chains. DeeperForward computes local cross-entropy from pooled goodness, but sends the class-associated feature maps to the next block. Anchor detaches the selected unsupported scores in the loss calculation; it does not detach or replace those propagated maps. HCL-FF applies the intervention through its routed cross-entropy while retaining its local contrastive components. The three branches therefore exercise the claimed unsupported-class setting and loss-side stop-gradient in the actual research architectures. The source manifest records the files used by these adapters, so the branch definitions can be checked against the native calls instead of inferred from a method name.
From a clean extraction, both full-width 17-layer, 100-class implementations completed all three branches for three epochs of two batches on eight real CIFAR-100 training examples. Every local optimizer completed six updates; states remained finite, and saved model checkpoints reproduced native predictions exactly after reloading. The command is python native_three_arm_small_training.py --kind DF --data-root DATA_ROOT --out DF_small.json --checkpoint-dir LOCAL_SMOKE_CHECKPOINTS; replace DF with HCL for the second implementation.
The updated package, 39008_Reviewer_Code_Materials.zip, also includes experiment configurations, run instructions, saved results, and result-recomputation scripts. Its README explains the entry points, and the reviewer guide links each concern to the relevant implementation and records.
Changes made: Updated the supporting materials with complete DF/HCL method implementations, experiment configurations, run instructions, and result-recomputation scripts.
Response 3: W5, W6 — Full-pool experiments and BP controls
W5 | Full training pool and larger settings
The added CIFAR-100 study uses all 50,000 official training images, split into 48,000 gradient-training and 2,000 validation images. Four systems and five seeds cover ten tasks; all 190 training units precede unified test evaluation. Anchor-minus-Matched mean arriving accuracy is +18.451 pp, while Anchor-minus-M0 final accuracy is -0.072 pp. These are the two prespecified paired comparisons; their 97.5% intervals are reported in Table 19. Final old-class accuracy remains below 1% for all systems. Tiny-ImageNet is mixed: the preset seed-0 gain is +4.8 points, below the preset 5-point criterion. Food101 has no Anchor arm.
This is a complete training-pool check, with validation separated from gradient updates. Each of the 100 classes contributes 480 training and 20 validation images. The same four systems follow ten ten-class tasks, using the recipes selected in the earlier two-task comparison without retuning on the full-pool test results. At each task boundary, prediction takes argmax over all seen classes: “current” and “old” identify subsets of test examples, not a smaller allowed label set. For each seed, arriving accuracy first averages tasks 2–10; paired uncertainty then uses the five seed-level differences. The 190 units comprise shared first-task prefixes and subsequent task training, yielding 20 system-by-seed trajectories, not 190 independent replications. This preserves the paired acquisition comparison while making its data coverage and statistical unit explicit.
Table R3. Full-pool CIFAR-100 results
| System | Arriving (%) | Final (%) | Old 90 (%) |
|---|---|---|---|
| Local M0 | 71.858 ± 0.253 | 8.502 ± 0.175 | 0.718 |
| Local Matched | 52.582 ± 1.404 | 8.484 ± 0.399 | 0.927 |
| Local Anchor | 71.033 ± 0.666 | 8.430 ± 0.299 | 0.738 |
| Terminal BP-M0 | 79.027 ± 0.612 | 8.984 ± 0.188 | 0.758 |
48,000 training and 2,000 validation images; ten tasks and five seeds. Arriving accuracy averages tasks 2–10. Arriving and final values are mean ± sample SD; Old 90 is the final mean. Sources: Section 5.6, Table 3 and Appendix Table 19.
Changes made: Added the full-pool study and paired results, and reported the exploratory Tiny-ImageNet and Food101 scope (Section 5.6; Appendix I.3, I.6–I.7; Table 19; Figure S11).
W6 | Beyond local learning
We added terminal-only and native-auxiliary-loss BP interventions. Current Anchor-minus-Matched effects are -0.500 and -0.520 pp, respectively. The full-pool terminal BP-M0 baseline also has higher arriving/final means than local Anchor. The common-prefix cut experiment then controls the starting state, recipe, losses, and readout; its interaction is -2.280 pp.
We separate a cross-system comparison from a controlled change in gradient connectivity. Terminal BP uses a final classifier; native-auxiliary BP sums all 17 local losses and predicts from the 17 native scores. In the cut tests, the common engine sums those losses, performs one backward pass and steps all 17 optimizers in both Cut and Coupled. Cutting block-input gradients leaves the forward feature maps and residual paths present. The 2×2 comparison then asks whether this backward change alters the Anchor-minus-Matched effect.
The new follow-up restores five native-local M0 first-task states and runs 20 fixed continuations in one common summed-loss engine. Cut and Coupled differ only in block-input detachment, with optimizer moments, replay and input/dropout streams preserved. Both prefix studies use the same constant-rate second-task recipe, which differs from the original native sequential-update experiment. Current accuracy is 26.46/37.16% for Cut Matched/Anchor and 30.92/27.52% for Coupled Matched/Anchor, always predicting among all 20 seen classes. The Cut effect is +10.700 pp; the Cut-minus-Coupled interaction is +14.100 pp. All 20 final states were locked before testing. Backward cuts strengthen the relative Anchor acquisition effect under this native-local starting configuration and continuation recipe. All four arms and old/current/final/AIA endpoints are reported.
Changes made: Added the native-local-prefix control while retaining the terminal BP, auxiliary BP, and auxiliary-prefix results (Section 5.7; Appendix I.2, I.4–I.5 and I.8; Figures 7 and S11).
Response 1: W1, W2 — Practical relevance and class allocation
Opening
Thank you for your detailed review and for identifying where stronger evidence and clearer documentation were needed. Your comments motivated experiments using the full training pool and further controls on the training configuration. We also revisited the research code and clarified the implementation details. Below, we respond to your comments individually and explain the corresponding revisions.
Added theory and experiments
The full-pool study uses 48,000 gradient-training and 2,000 validation images. Terminal and native-auxiliary BP controls and a common-prefix study test the limits of the local result (Sections 5.6–5.7; Appendix I). We separate the depth cohorts rather than pooling different exposure histories (Section 5.2; Appendix D.2–D.3). A fixed 20-continuation follow-up from five native-local prefixes adds a positive conditional cut interaction.
We have updated the supporting materials with the complete code for our method. Experiment configurations and result-recomputation scripts are also included.
Figure changes
Figure 3(b) and Figure S5 separate the depth and phase cohorts. Figure S10 removes the pooled heterogeneous depth estimate and uses the original three-seed current-versus-neither comparison. Added Figure S11 displays five prespecified paired contrasts from the full-pool, auxiliary-BP, and common-prefix studies, with their distinct endpoints and intervals. Figure 7 in main-text Section 5.7 shows the new native-local-prefix seed pairs and its two contrasts prespecified in this follow-up.
W1 | Avoidability and practical relevance
Unsupported outputs can be masked. The experiment asks a narrower question: can acquisition recover while numerical denominator participation remains? The common-backbone local comparison gives 17.10/43.60/44.72% current accuracy for Matched/Anchor/M0, and the full-pool comparison gives an 18.451-point arriving gain over Matched. These results identify acquisition recovery while retaining denominator participation.
The architectural motivation is that class-associated feature maps serve two roles: they are locally supervised, and they become inputs to later blocks. Their forward role persists even when the corresponding class has not yet supplied positive examples. This makes the interaction between local supervision and later representation learning a concrete question. Our comparison keeps the feature bank and its forward transmission, and changes how unsupported coordinates participate in the local loss. It shows an avoidable acquisition cost in the tested configuration and a way to isolate its derivative contribution.
Changes made: Stated the diagnostic question tested by Anchor (Introduction; Section 6).
W2 | Unknown class count
Our experiments assume a known preallocated output capacity. Expanding class-associated feature groups can also change the next block's input width; masking a fixed bank is not dynamic expansion. The full-pool study still assumes 100-class capacity.
Changes made: Specified the known-capacity assumption and distinguished fixed-bank masking from dynamic expansion (Section 2; Section 6).